Scholarlink

<p>It is challenging to establish accurate and efficient surface quality measurements in finishing processes due to numerous influencing parameters involved. This study aims to develop an image-based classification model for assessing surface improvement of Ni-Ti alloy in a rotational electro-magnetic finishing process using the EfficientNet-B0 algorithm as a novel approach within artificial neural networks. To improve classification accuracy and evaluate the relationship between abrasive motion and surface improvement, trained and test images that indicated the degree of surface finishing performance through simulations were classified into five categories based on three criteria: contact time, contact distribution, and the magnitude of impact force. The EfficientNet-B0 algorithm achieved accuracies of 98.4 % and 94.1 % for training and testing data, respectively, which outperformed the traditional convolutional neural network that achieved 91.7 % and 83.3 % for the same datasets. Both simulated and predicted results demonstrated reliability and generalization capabilities, consistently aligning with experimental verification.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Scaling efficient net based on structural transient analysis for multi-class classification of Ni-Ti alloy surface improvement

  • 期刊论文

  • 发表日期：

    2025年10月8日

查看全文
  • Jung-Hee Lee,
  • Yeo-Kyung Jung,
  • Jae-Seob Kwak

摘要

It is challenging to establish accurate and efficient surface quality measurements in finishing processes due to numerous influencing parameters involved. This study aims to develop an image-based classification model for assessing surface improvement of Ni-Ti alloy in a rotational electro-magnetic finishing process using the EfficientNet-B0 algorithm as a novel approach within artificial neural networks. To improve classification accuracy and evaluate the relationship between abrasive motion and surface improvement, trained and test images that indicated the degree of surface finishing performance through simulations were classified into five categories based on three criteria: contact time, contact distribution, and the magnitude of impact force. The EfficientNet-B0 algorithm achieved accuracies of 98.4 % and 94.1 % for training and testing data, respectively, which outperformed the traditional convolutional neural network that achieved 91.7 % and 83.3 % for the same datasets. Both simulated and predicted results demonstrated reliability and generalization capabilities, consistently aligning with experimental verification.